Wednesday, July 3, 2019

Using AI to solve data and IT complexity -- and thereby better enable AI


The next BriefingsDirect data disruption discussion focuses on why the rising tidal wave of data must be better managed, and how new tools are emerging to bring artificial intelligence (AI) to the rescue. 

Stay with us to explore how the latest AI innovations improve both data and services management across a cloud deployment continuum -- and in doing so set up an even more powerful way for businesses to exploit AI.

Listen to the podcast. Find it on iTunes. Read a full transcript or download a copy.

To learn how AI will help conquer complexity to allow for higher abstractions of benefits from across all sorts of analysis, we welcome Rebecca Lewington, Senior Manager of Innovation Marketing at Hewlett Packard Enterprise (HPE). The interview is conducted by Dana Gardner, Principal Analyst at Interarbor Solutions.

Here are some excerpts:

Gardner: We have been talking about massive amounts of data for quite some time. What’s new about data buildup that requires us to look to AI for help?

Lewington: Partly it is the sheer amount of data. IDC’s Data Age Study predicts the global data sphere will be 175 zettabytes by 2025, which is a rather large number. That’s what, 1 and 21 zeros? But we have always been in an era of exploding data.

Lewington
Yet, things are different. One, it’s not just the amount of data; it’s the number of sources the data comes from. We are adding in things like mobile devices, and we are connecting factories’ operational technologies to information technology (IT). There are more and more sources.

Also, the time we have to do something with that data is shrinking to the point where we expect everything to be real-time or you are going to make a bad decision. An autonomous car, for example, might do something bad. Or we are going to miss a market or competitive intelligence opportunity.

So it’s not just the amount of data -- but what you need to do with it that is challenging.

Gardner: We are also at a time when Al and machine learning (ML) technologies have matured. We can begin to turn them toward the data issue to better exploit the data. What is new and interesting about AI and ML that make them more applicable for this data complexity issue?

Data gets smarter with AI

Lewington: A lot of the key algorithms for AI were actually invented long ago in the 1950s, but at that time, the computers were hopeless relative to what we have today; so it wasn’t possible to harness them.

For example, you can train a deep-learning neural net to recognize pictures of kittens. To do that, you need to run millions of images to train a working model you can deploy. That’s a huge, computationally intensive task that only became practical a few years ago. But now that we have hit that inflection point, things are just taking off.

Gardner: We can begin to use machines to better manage data that we can then apply to machines. Does that change the definition of AI?

Lewington: The definition of AI is tricky. It’s malleable, depending on who you talk to. For some people, it’s anything that a human can do. To others, it means sophisticated techniques, like reinforcement learning and deep learning.
How to Remove Complexity
From Multicloud and Hybrid IT
One useful definition is that AI is what you use when you know what the answer looks like, but not how to get there.

Traditional analytics effectively does at scale what you could do with pencil and paper. You could write the equations to decide where your data should live, depending on how quickly you need to access it.

But with AI, it’s like the kittens example. You know what the answer looks like, it’s trivial for you to look at the photograph and say, “That is a cat in the picture.” But it’s really, really difficult to write the equations to do it. But now, it’s become relatively easy to train a black box model to do that job for you.

Gardner: Now that we are able to train the black box, how can we apply that in a practical way to the business problem that we discussed at the outset? What is it about AI now that helps better manage data? What's changed that gives us better data because we are using AI?
The heart of what makes AI work is good data; the right data, in the right place, with the right properties you can use to train a model, which you can then feed new data into to get results that you couldn't get otherwise.

Lewington: It’s a circular thing. The heart of what makes AI work is good data; the right data, in the right place, with the right properties you can use to train a model, which you can then feed new data into to get results that you couldn’t get otherwise.

Now, there are many ways you can apply that. You can apply it to the trivial case of the cat we just talked about. You can apply it to helping a surgeon review many more MRIs, for example, by allowing him to focus on the few that are borderline, and to do the mundane stuff for him.

But, one of the other things you can do with it is use it to manipulate the data itself. So we are using AI to make the data better -- to make AI better.

Gardner: Not only is it circular, and potentially highly reinforcing, but when we apply this to operations in IT -- particularly complexity in hybrid cloud, multicloud, and hybrid IT -- we get an additional benefit. You can make the IT systems more powerful when it comes to the application of that circular capability -- of making better AI and better data management.

AI scales data upward and outward

Lewington: Oh, absolutely. I think the key word here is scale. When you think about data -- and all of the places it can be, all the formats it can be in -- you could do it yourself. If you want to do a particular task, you could do what has traditionally been done. You can say, “Well, I need to import the data from here to here and to spin up these clusters and install these applications.” Those are all things you could do manually, and you can do them for one-off things.

But once you get to a certain scale, you need to do them hundreds of times, thousands of times, even millions of times. And you don’t have the humans to do it. It’s ridiculous. So AI gives you a way to augment the humans you do have, to take the mundane stuff away, so they can get straight to what they want to do, which is coming up with an answer instead of spending weeks and months preparing to start to work out the answer.

Gardner: So AI directed at IT, what some people call AIOps could be an accelerant to this circular advantageous relationship between AI and data? And is that part of what you are doing within the innovation and research work at HPE?

Lewington: That’s true, absolutely. The mission of Hewlett Packard Labs in this space is to assist the rest of the company to create more powerful, more flexible, more secure, and more efficient computing and data architectures. And for us in Labs, this tends to be a fairly specific series of research projects that feed into the bigger picture.

https://www.hpe.com/us/en/resources/solutions/deep-learning-dummies-gen10.html?chatsrc=ot-en&jumpid=ps_17fix8scuz_aid-510455007&gclid=EAIaIQobChMIp-OTod3k4gIVFqSzCh2rUwd0EAAYASAAEgIC3fD_BwE&gclsrc=aw.ds

For example, we are now doing the Deep Learning Cookbook, which allows customers to find out ahead of time exactly what kind of hardware and software they are going to need to get to a desired outcome. We are automating the experimenting process, if you will.

And, as we talked about earlier, there is the shift to the edge. As we make more and more decisions -- and gain more insights there, to where the data is created -- there is a growing need to deploy AI at the edge. That means you need a data strategy to get the data in the right place together with the AI algorithm, at the edge. That’s because there often isn’t time to move that data into the cloud before making a decision and waiting for the required action to return.

Once you begin doing that, once you start moving from a few clouds to thousands and millions of endpoints, how do you handle multiple deployments? How do you maintain security and data integrity across all of those devices? As researchers, we aim to answer exactly those questions.

And, further out, we are looking to move the natural learning phase itself to the edge, to do the things we call swarm learning, where devices learn from their environment and each other, using a distributed model that doesn’t use a central cloud at all.

Gardner: Rebecca, given your title is Innovation Marketing Lead, is there something about the very nature of innovation that you have come to learn personally that’s different than what you expected? How has innovation itself changed in the past several years?

Innovation takes time and space 

Lewington: I began my career as a mechanical engineer. For many years, I was offended by the term innovation process, because that’s not how innovation works. You give people the space and you give them the time and ideas appear organically. You can’t have a process to have ideas. You can have a process to put those ideas into reality, to wean out the ones that aren’t going to succeed, and to promote the ones that work.
How to Better Understand
What AI Can do For Your Business
But the term innovation process to me is an oxymoron. And that’s the beautiful thing about Hewlett Packard Labs. It was set up to give people the space where they can work on things that just seem like a good idea when they pop up in their heads. They can work on these and figure out which ones will be of use to the broader organization -- and then it’s full steam ahead.

Gardner: It seems to me that the relationship between infrastructure and AI has changed. It wasn’t that long ago when we thought of business intelligence (BI) as an application -- above the infrastructure. But the way you are describing the requirements of management in an edge environment -- of being able to harness complexity across multiple clouds and the edge -- this is much more of a function of the capability of the infrastructure, too. Is that how you are seeing it, that only a supplier that’s deep in its infrastructure roots can solve these problems? This is not a bolt-on benefit.

Lewington: I wouldn’t say it’s impossible as a bolt-on; it’s impossible to do efficiently and securely as a bolt-on. One of the problems with AI is we are going to use a black box; you don’t know how it works. There were a number of news stories recently about AIs becoming corrupted, biased, and even racist, for example. Those kinds of problems are going to become more common.

And so you need to know that your systems maintain their integrity and are not able to be breached by bad actors. If you are just working on the very top layers of the software, it’s going to be very difficult to attest that what’s underneath has its integrity unviolated.

If you are someone like HPE, which has its fingers in lots of pies, either directly or through our partners, it’s easier to make a more efficient solution.
You need to know that your systems maintain their integrity and are not able to be breached by bad actors. If you are just working on the very top layers of the software, it's going to be very difficult to attest that what's underneath has its integrity unviolated.

Gardner: Is it fair to say that AI should be a new core competency, for not only data scientists and IT operators, but pretty much anybody in business? It seems to me this is an essential core competency across the board.

Lewington: I think that's true. Think of AI as another layer of tools that, as we go forward, becomes increasingly sophisticated. We will add more and more tools to our AI toolbox. And this is one set of tools that you just cannot afford not to have.

Gardner: Rebecca, it seems to me that there is virtually nothing within an enterprise that won't be impacted in one way or another by AI.

Lewington: I think that’s true. Anywhere in our lives where there is an equation, there could be AI. There is so much data coming from so many sources. Many things are now overwhelmed by the amount of data, even if it’s just as mundane as deciding what to read in the morning or what route to take to work, let alone how to manage my enterprise IT infrastructure. All things that are rule-based can be made more powerful, more flexible, and more responsive using AI.

Gardner: Returning to the circular nature of using AI to make more data available for AI -- and recognizing that the IT infrastructure is a big part of that -- what are doing in your research and development to make data services available and secure? Is there a relationship between things like HPE OneView and HPE OneSphere and AI when it comes to efficiency and security at scale?

Let the system deal with IT 

Lewington: Those tools historically have been rules-based. We know that if a storage disk gets to a certain percentage full, we need to spin up another disk -- those kinds of things. But to scale flexibly, at some point that rules-based approach becomes unworkable. You want to have the system look after itself, to identify its own problems and deal with them.

Including AI techniques in things like HPE InfoSight, HPE Clearpath, and network user identity behavior software on the HPE Aruba side allows the AI algorithms to make those tools more powerful and more efficient.

https://www.hpe.com/us/en/resources/solutions/deep-learning-dummies-gen10.html?chatsrc=ot-en&jumpid=ps_17fix8scuz_aid-510455007&gclid=EAIaIQobChMIp-OTod3k4gIVFqSzCh2rUwd0EAAYASAAEgIC3fD_BwE&gclsrc=aw.ds
You can think of AI here as another class of analytics tools. It’s not magic, it’s just a different and better way of doing IT analytics. The AI lets you harness more difficult datasets, more complicated datasets, and more distributed datasets.

Gardner: If I’m an IT operator in a global 2000 enterprise, and I’m using analytics to help run my IT systems, what should I be thinking about differently to begin using AI -- rather than just analytics alone -- to do my job better?

Lewington: If you are that person, you don’t really want to think about the AI. You don’t want the AI to intrude upon your consciousness. You just want the tools to do your job.

For example, I may have 1,000 people starting a factory in Azerbaijan, or somewhere, and I need to provision for all of that. I want to be able to put on my headset and say, “Hey, computer, set up all the stuff I need in Azerbaijan.” You don’t want to think about what’s under the hood. Our job is to make those tools invisible and powerful.

Composable, invisible, and insightful 

Gardner: That sounds a lot like composability. Is that another tangent that HPE is working on that aligns well with AI?

Lewington: It would be difficult to have AI be part of the fabric of an enterprise without composability, and without extending composability into more dimensions. It’s not just about being able to define the amount of storage and computer networking with a line of code, it’s about being able to define the amount of memory, where the data is, where the data should be, and what format the data should be in. All of those things – from the edge to cloud – need to be dimensions in composability.
How to Achieve Composability
Across Your Datacenter
You want everything to work behind the scenes for you in the best way with the quickest results, with the least energy, and in the most cost-effective way possible. That’s what we want to achieve -- invisible infrastructure.

Gardner: We have been speaking at a fairly abstract level, but let’s look to some examples to illustrate what we’re getting at when we think about such composability sophistication.

Do you have any concrete examples or use cases within HPE that illustrate the business practicality of what we’ve been talking about?


Lewington: Yes, we have helped a tremendous number of customers either get started with AI in their operations or move from pilot to volume use. A couple of them stand out. One particular manufacturing company makes electronic components. They needed to improve the yields in their production lines, and they didn’t know how to attack the problem. We were able to partner with them to use such things as vision systems and photographs from their production tools to identify defects that only could be picked up by a human if they had a whole lot of humans watching everything all of the time.

This gets back to the notion of augmenting human capabilities. Their machines produce terabytes of data every day, and it just gets turned away. They don’t know what to do with it.

We began running some research projects with them to use some very sophisticated techniques, visual autoencoders, that allow you, without having a training set, to characterize a production line that is performing well versus one that is on the verge of moving away from the sweet spot. Those techniques can fingerprint a good line and also identify when the lines go just slightly bad. In that case, a human looking at line would think it was working just perfectly.

This takes the idea of predictive maintenance further into what we call prescriptive maintenance, where we have a much more sophisticated view into what represents a good line and what represents a bad line. Those are couple of examples for manufacturing that I think are relevant.

Gardner: If I am an IT strategist, a Chief Information Officer (CIO) or a Chief Technology Officer (CTO), for example, and I’m looking at what HPE is doing -- perhaps at the HPE Discover conference -- where should I focus my attention if I want to become better at using AI, even if it’s invisible? How can I become more capable as an organization to enable AI to become a bigger part of what we do as a company?

The new company man is AI

Lewington: For CIOs, their most important customers these days may be developers and increasingly data scientists, who are basically developers working with training models as opposed to programs and code. They don’t want to have to think about where that data is coming from and what it’s running on. They just want to be able to experiment, to put together frameworks that turn data into insights.

It’s very much like the programming world, where we’ve gradually abstracted things from bare-metal, to virtual machines, to containers, and now to the emerging paradigm of serverless in some of the walled-garden public clouds. Now, you want to do the same thing for that data scientist, in an analogous way.

https://www.hpe.com/us/en/solutions/cloud/composable-private-cloud.html

Today, it’s a lot of heavy lifting, getting these things ready. It’s very difficult for a data scientist to experiment. They know what they want. They ask for it, but it takes weeks and months to set up a system so they can do that one experiment. Then they find it doesn’t work and move on to do something different. And that requires a complete re-spin of what’s under the hood.

Now, using things like software from the recent HPE BlueData acquisition, we can make all of that go away. And so the CIO’s job becomes much simpler because they can provide their customers the tools they need to get their work done without them calling up every 10 seconds and saying, “I need a cluster, I need a cluster, I need a cluster.”

That’s what a CIO should be looking for, a partner that can help them abstract complexity away, get it done at scale, and in a way that they can both afford and that takes the risk out. This is complicated, it’s daunting, and the field is changing so fast.

Gardner: So, in a nutshell, they need to look to the innovation that organizations like HPE are doing in order to then promulgate more innovation themselves within their own organization. It’s an interesting time.

Containers contend for the future 

Lewington: Yes, that’s very well put. Because it’s changing so fast they don’t just want a partner who has the stuff they need today, even if they don’t necessarily know what they need today. They want to know that the partner they are working with is working on what they are going to need five to 10 years down the line -- and thinking even further out. So I think that’s one of the things that we bring to the table that others can’t.

Gardner: Can give us a hint as to what some of those innovations four or five years out might be? How should we not limit ourselves in our thinking when it comes to that relationship, that circular relationship between AI, data, and innovation?

Lewington: It was worth coming to HPE Discover in June, because we talked about some exciting new things around many different options. The discussion about increasing automation abstractions is just going to accelerate.
We are going to get to the point where using containers seems as complicated as bare-metal today and that's really going to help simplify the whole data pipelines thing.

For example, the use of containers, which have a fairly small penetration rate across enterprises, is at about 10 percent adoption today because they are not the simplest thing in the world. But we are going to get to the point where using containers seems as complicated as bare-metal today and that’s really going to help simplify the whole data pipelines thing.

Beyond that, the elephant in the room for AI is that model complexity is growing incredibly fast. The compute requirements are going up, something like 10 times faster than Moore’s Law, even as Moore’s Law is slowing down.

We are already seeing an AI compute gap between what we can achieve and what we need to achieve -- and it’s not just compute, it’s also energy. The world’s energy supply is going up, can only go up slowly, but if we have exponentially more data, exponentially more compute, exponentially more energy, and that’s just not going to be sustainable.

So we are also working on something called Emergent Computing, a super-energy-efficient architecture that moves data around wherever it needs to be -- or not move data around but instead bring the compute to the data. That will help us close that gap.
How to Transform
The Traditional Datacenter
And that includes some very exciting new accelerator technologies: special-purpose compute engines designed specifically for certain AI algorithms. Not only are we using regular transistor-logic, we are using analog computing, and even optical computing to do some of these tasks, yet hundreds of times more efficiently and using hundreds of times less energy. This is all very exciting stuff, for a little further out in the future.

No comments:

Post a Comment